Review Text, Find Errors, and Check Unsupported Information with AI
Use a structured prompt to edit language and logic, identify claims needing evidence, and verify sources with a workflow that avoids false confidence.
Separate the kinds of review
Run spelling, grammar, and style review first; logic and consistency second; and claim verification as a separate task. Combining them can make polished wording look like proof that the content is true.
Use: “Edit this text for language without changing meaning. Show the original, revision, and reason; do not rewrite the entire text.” Then run: “Extract every verifiable claim, identify what needs a source, and do not guess a source.”
Finding unsupported information
Create a table with the claim, type, cited source, clarity, and evidence required. Pay special attention to numbers, dates, quotations, and causal language; they may sound plausible while requiring an original record or study.
Tell the model to mark an item “unverified” rather than fill a gap. Check the primary source or official document yourself and confirm it supports the same claim in context. If verification fails, qualify or remove it instead of presenting it as fact.
A repeatable comprehensive review prompt
Try: “You are a meticulous editor. Review for language errors, ambiguous terms, internal contradictions, claims needing evidence, and improvements. For each note, quote the sentence, classify the issue, explain it, and suggest a meaning-preserving revision.”
Include audience, purpose, and style rules, and ask to preserve the writer’s voice. Do not remove caveats or sources merely for smoother prose. Read the complete final version afterward because a local edit can conflict with a distant paragraph.
Final verification and privacy
Sort results into confirmed error, style suggestion, and claim requiring verification. Manually review decisions, quotations, numbers, and links, and record source access dates. AI can flag a questionable passage, but it does not prove a fact or guarantee a complete review.
Remove customer data and confidential information before entering an unpublished draft into a public tool, and retain the original and change log in an authorized location. A safe workflow is redact, edit, extract claims, verify independently, review, then publish.